AgentSkillsCN

deep-research

制定详尽、递归深入的标准化操作流程。通过事实核查、递归爬取与多维度搜索等方法,全面梳理与呈现主题内容。

SKILL.md
--- frontmatter
name: deep-research
description: Standard operating procedure for exhaustive, recursive deep-dives. Provides protocols for fact-checking, recursive crawling, and multi-vector search to map topics completely.

Research Protocol

  1. Autonomous Deep-Dive: Explore until the topic is fully mapped. Continue until all questions are answered.
  2. Zero Assumptions: Verify everything. Find the primary source for every claim. Cross-reference niche data.
  3. Recursive Crawling: Follow every relevant link. If a page has sub-links, crawl them all.
  4. Multi-Vector Search: Prioritize the websearch tool (Exa AI) for deep, real-time information retrieval. Use 5-10 query variations. If websearch or codesearch yields insufficient results, switch strategies immediately.
  5. Bypass Blockers: If webfetch is restricted or content is hidden, use browser_execute to scroll, click, and interact.
  6. Iterative Note-Taking: Research → Discover → Write Note → Repeat. Capture insights while fresh.

Execution Checklist

  • Crawl everything: No link left unclicked.
  • Search everywhere: Exhaust all tool vectors and query variations using Exa (websearch).
  • Verify [?]: Immediately prioritize and resolve uncertain findings.
  • Source URLs: Mandatory for every finding. Use archives/cache if blocked.
  • No stop: Don't check in or pause until the topic is exhausted.

Note Format

Use YAML frontmatter with tags. Every entry must be timestamped:

markdown
### [YYYY-MM-DD HH:MM] Finding Title

**Source**: <url>
**Key Points**:

- Point 1
- Point 2
  **Relevance**: <connection to goal>

---

Termination Criteria

Research is complete only when:

  1. Every relevant link in every source has been followed.
  2. New search queries return no new information.
  3. All [?] marks are resolved and verified.
  4. Every logical "how" and "why" has a documented source.